DX11（SharpDX）计算着色器结果读取异常排查求助

我用SharpDX开发DirectX 11计算着色器，预期输出应为{2,2,2}, {5,5,5}, {8,8,8}，但实际得到{2,5,8}, {0,0,0}, {0,0,0}；将HLSL输出改为xyz时，结果变成{1,4,7}, {0,0,0}, {0,0,0}。确认计算着色器输入和逻辑没问题，问题应该出在GPU结果读取环节，求排查。

C#代码实现 #

float[] buffer = new float[]{1,2,3,4,5,6,7,8,9};
var vertexArray = MemoryMarshal.Cast<float, Vector3>(buffer).ToArray();
var device = new Device(DriverType.Hardware, DeviceCreationFlags.Debug);
var context = device.ImmediateContext;
// 编译计算着色器
var faceCsFile = Path.Combine(AppContext.BaseDirectory, "HLSL/Face_CS.hlsl");
var shaderBytecode = ShaderBytecode.CompileFromFile(faceCsFile, "CS", "cs_5_0");
var computeShader = new ComputeShader(device, shaderBytecode);
shaderBytecode.Dispose();
var inputDesc = new BufferDescription()
{
    SizeInBytes = SharpDX.Utilities.SizeOf<Vector3>() * vertexArray.Length,
    Usage = ResourceUsage.Default,
    BindFlags = BindFlags.ShaderResource | BindFlags.UnorderedAccess,
    OptionFlags = ResourceOptionFlags.BufferStructured,
    StructureByteStride = SharpDX.Utilities.SizeOf<Vector3>(),
};
var verticeBuffer = Buffer.Create(device, vertexArray, inputDesc);
device.ImmediateContext.UnmapSubresource(verticeBuffer, 0);
var srvDesc = new ShaderResourceViewDescription()
{
    Format = SharpDX.DXGI.Format.Unknown,
    Dimension = ShaderResourceViewDimension.Buffer,
    BufferEx = new ShaderResourceViewDescription.ExtendedBufferResource()
    {
        FirstElement = 0,
        ElementCount = vertexArray.Length,
        Flags = ShaderResourceViewExtendedBufferFlags.None
    } 
};
var vertexBufferView = new ShaderResourceView(device, verticeBuffer, srvDesc);
context.ComputeShader.SetShaderResource(0, vertexBufferView);
// 将输入和输出缓冲区绑定到计算着色器上
context.ComputeShader.Set(computeShader);
var dataOut = new Vector3[vertexArray.Length];
var outDes = new BufferDescription()
{
    SizeInBytes = SharpDX.Utilities.SizeOf<Vector3>() * vertexArray.Length,
    Usage = ResourceUsage.Default,
    BindFlags = BindFlags.UnorderedAccess,
    OptionFlags = ResourceOptionFlags.BufferStructured,
    StructureByteStride = SharpDX.Utilities.SizeOf<Vector3>(),
};
var bufferOut = Buffer.Create(device, dataOut, outDes);
// 创建用于指定着色器输出的UAV
var uavDesc = new UnorderedAccessViewDescription()
{
    Format = SharpDX.DXGI.Format.Unknown,
    Dimension = UnorderedAccessViewDimension.Buffer,
    Buffer = new UnorderedAccessViewDescription.BufferResource()
    {
        ElementCount = vertexArray.Length,
    }
};
var bufferOutView = new UnorderedAccessView(device, bufferOut, uavDesc);
context.ComputeShader.SetUnorderedAccessView(0, bufferOutView);
int count = (vertexArray.Length + 1023) / 1024;
context.Dispatch(count, 1, 1);// 启动计算着色器
var stagingBuffer = new Buffer(device, new BufferDescription
{
    BindFlags = BindFlags.None,
    CpuAccessFlags = CpuAccessFlags.Read,
    OptionFlags = ResourceOptionFlags.None,
    SizeInBytes = dataOut.Length * SharpDX.Utilities.SizeOf<Vector3>(),
    Usage = ResourceUsage.Staging
});// 创建一个暂存缓冲区
context.CopyResource(bufferOut, stagingBuffer);// 将输出缓冲区的内容复制到暂存缓冲区中
// 将暂存缓冲区映射到内存中
var dataBox = context.MapSubresource(stagingBuffer, 0, MapMode.Read, SharpDX.Direct3D11.MapFlags.None);
Span<Vector3> result;
unsafe{result = new Span<Vector3>((void*)dataBox.DataPointer, dataOut.Length);}// 读取数据
context.UnmapSubresource(stagingBuffer, 0);// 取消映射
foreach (var vec in result.ToArray())
    Console.WriteLine(vec);

HLSL计算着色器代码 #

StructuredBuffer<float3> Vertex : register(t0);
RWBuffer<float3> Out : register(u0);

[numthreads(1024, 1, 1)]
void CS(uint3 DTid : SV_DispatchThreadID)
{
    uint id = DTid.x;
    Out[id]=Vertex[id].yyy;
}

问题根源分析 #

1. 暂存缓冲区属性不匹配
   输出缓冲区是结构化缓冲区（ResourceOptionFlags.BufferStructured），但创建的暂存缓冲区未指定结构化属性，导致CPU映射时无法正确解析内存布局，把连续的float数据错误地按Vector3结构读取。

2. GPU计算与CPU复制不同步
   Dispatch调用后GPU可能还未完成计算就执行CopyResource，导致复制的是未初始化或不完整的数据。

修正方案 #

1. 修正暂存缓冲区创建
   让暂存缓冲区和输出缓冲区保持一致的结构化属性：

   var stagingBuffer = new Buffer(device, new BufferDescription
   {
       BindFlags = BindFlags.None,
       CpuAccessFlags = CpuAccessFlags.Read,
       OptionFlags = ResourceOptionFlags.BufferStructured,
       StructureByteStride = SharpDX.Utilities.SizeOf<Vector3>(),
       SizeInBytes = dataOut.Length * SharpDX.Utilities.SizeOf<Vector3>(),
       Usage = ResourceUsage.Staging
   });
   

2. 添加资源同步屏障
   在Dispatch后、CopyResource前插入屏障，确保计算完成后再复制数据：

   context.Dispatch(count, 1, 1);
   // 添加资源屏障同步GPU计算
   var barrier = new ResourceBarrier(
       ResourceBarrierType.Transition,
       ResourceBarrierFlags.None,
       bufferOut,
       ResourceState.UnorderedAccess,
       ResourceState.CopySource
   );
   context.ResourceBarrier(barrier);
   // 执行复制操作
   context.CopyResource(bufferOut, stagingBuffer);
   

3. 移除多余代码
   删除device.ImmediateContext.UnmapSubresource(verticeBuffer, 0);，Buffer.Create已经完成资源初始化，无需手动Unmap。

内容的提问来源于stack exchange，提问作者ThraceShah